ICP Evidence Analysis

Build hypotheses from evidence. Do not write persona fiction.

This skill inherits the ethical floor from skills/ETHICS.md. It enforces anti-hallucination and permission checks. Anti-spray and human-send are not applicable because this skill produces research artifacts, not outreach.

Inputs

Accept any combination of:

• a company URL and user-supplied description;
• source_manifest.json;
• public product, pricing, integration, security, support, certification, filing, review, or coverage pages;
• an existing Newsjack monitor profile as unverified leads;
• target markets, exclusions, permitted-data rules, or fact-check results.

If only a URL and description are supplied, research enough independent public evidence to test the website's claims. Never treat a target's own positioning as buyer behavior.

Build the evidence perimeter

For every material claim, retain:

• source ID, canonical URL or authorized file;
• title/publisher, access or publication date, and source type;
• exact span or a faithful short paraphrase;
• whether it is company_asserted, buyer_behavior, or independent;
• fact type, confidence, and permission status.

Keep negative and conflicting evidence. Stop if a source is private, unlawfully obtained, outside the user's permission, or contains personal data that is not needed.

Separate:

• products, capabilities, limitations, integrations, proof assets, and geographies;
• declared segments from independently supported demand;
• exact brand, product, domain, people, slogan, proprietary-category, campaign, and flattering-claim terms for the later contamination register;
• competitors named by evidence from alternatives merely guessed by the model.

Classify by publisher provenance, not by the claim's tone. A target-authored postmortem, benchmark, customer story, or technical article remains company_asserted; split it from any third-party analysis instead of blending both into one independent record.